janas-chat

janas-chat [model] [options]

janas-chat is the terminal client of Janas-LLM, a family of C programs that run mixture-of-experts language models whose weights stay on disk. Experts each token needs are streamed from an NVMe SSD; the ones that keep coming back stay in an in-memory cache. An 80-billion-parameter model can answer on a laptop with 32 GB of RAM.model is a .jns file, or a name from the catalog (`janas-get list` shows them). With no model, the chat lists the catalog, marks what is already on disk, and asks which one. A catalog model that is not in `~/.local/share/janas/models` (or JANAS_MODELS / --models) is downloaded and converted after asking, through janas-get.On a terminal the program takes the window: a banner, the conversation, and a footer with the model name, context used, last speed, and expert-cache fill. Lines starting with / are commands (`/help` lists them). A backslash then Enter adds a line to the message instead of sending it. Ctrl-C stops a reply; Ctrl-D quits. Piped in or out it stays a plain chat for scripts.Settings stack in this order: the model's catalog profile, the measure of the model on this machine, `~/.config/janas/chat.conf`, `~/.config/janas/models/<name>.conf`, then the command line. `/save-config` writes the model's file (`/save-config global` the chat's). The first start of a model on a machine can offer to measure it (`--tune` / `--no-tune`).A model whose catalog entry names a draft or multi-token prediction file gets it by itself. --mtp names one; --no-mtp goes without. --draft puts a small model of the same family in place of conversation-copied guesses (worth it on dense models). Speculative decoding does not change which tokens come out, only how many a pass checks; --no-spec and `/spec off` turn it off.

model

A `.jns` path or catalog name. None: list the catalog and ask.

Catalog profile. quality (default) keeps answers as the file gives them; fast adds levers measured to make that model faster at a small cost to its answers.

Where models live (default JANAS_MODELS, or `~/.local/share/janas/models`).

Measure again how the model runs fastest on this machine, or neither ask nor use that measure.

Multi-token prediction file (an MTP block, or Gemma 4's assistant). Without it, the matching file beside the model.

No prediction file, not even one found beside the model.

A small model of the same family to guess the next tokens.

Context length (default 16384, never more than the model was trained for).

Memory for streamed experts (default: what the machine can spare).

When the cache is automatic, memory left to other programs (default: a fifth of the machine's).

Bits per weight of the experts' down matrix, for files stored in planes (default: the memory decides).

Bits of the output head. 4 requantizes a six-bit head at start (default 6, as in the file).

How attention scores are computed (default: fast above 16384 tokens of context).

System message (`""` for none).

Sampling. Defaults follow Qwen's recommendations: temperature 0.7, top-k 20, top-p 0.8.

Longest reply (default: no limit).

Power mode. eco never uses the GPU for speed alone.

Reasoning before replying, for models that do it (default off).

A line of counters after every reply.

Never give the GPU work, whatever the mode says.

No speculative decoding.

Print the model's marks instead of rendering them.

MCP servers whose tools the model may call (default `~/.config/janas/mcp.json`).

Start no MCP server, including Janas's own services.

Start none of Janas's own services (`janas-flights` and the others beside the chat).

A line on stderr every s seconds while a long prompt is read.

~/.local/share/janas/models

Default model directory.

Global chat settings (`/save-config global`).

Per-model settings (`/save-config`).

Tuning profiles, expert-cache learning, and other machine-local state.

MCP servers in the usual `mcpServers` format.

Override the model directory.

Keep the engine off the GPU entirely.

Turn self-tuning off (fixed defaults).

Janas is early software: the public API and the command line change without notice. The engine fills free memory with its expert cache and reads the disk with O_DIRECT; on a machine with less headroom than it reckons, the desktop can be pushed into swap. Using a GPU has, once, caused a driver reset that took the desktop with it. It was developed on one Linux x86-64 laptop (Debian 13, Intel Core Ultra 9 185H, 32 GB, NVMe); other CPUs, discrete GPUs, slower disks, and smaller machines are untested. macOS is not supported. Models must be .jns files (converted from GGUF); IQ2 and IQ1 types are refused. Qwen3.5-0.8B is a test model left out of the pick list.

Janas is designed and written by Maurizio "camauri" Cammalleri. The first public release of Janas-LLM was in September 2026. It is free software under the GNU GPL version 3 or later.

janas-get(1), janas-bench(1), janas-server(1), janas-mcp(1), llama-cli(1), ollama(1), janus(1), koboldcpp(1)

Source code · Documentation